AMD introduces the Instinct MI450X with 288 GB of HBM4
AMD's newsroom post, published 12 November 2026 at 8:00 AM CST, introduces the Instinct MI450X GPU for AMD Helios. It says the MI450X offers 288 GB of HBM4 memory per GPU, with the same core compute, memory bandwidth, and scale-up and scale-out connectivity as the MI455X. It says customers can select either GPU on the same rack-scale architecture and ROCm software, and that the MI455X provides additional memory headroom.
Source: AMD (opens in a new tab), 12 Nov 2026
AMD's newsroom post, published 12 November 2026 at 8:00 AM CST, introduces the Instinct MI450X GPU for AMD Helios. It says the MI450X offers 288 GB of HBM4 memory per GPU, with the same core compute, memory bandwidth, and scale-up and scale-out connectivity as the MI455X. It says customers can select either GPU on the same rack-scale architecture and ROCm software, and that the MI455X provides additional memory headroom.
The same post says AMD testing of the MI455X on Helios delivered 21,333 tokens per second per GPU on DeepSeek-R1, using an MLPerf Inference Offline configuration, which it describes as up to 30% higher throughput than the Nvidia Vera Rubin latest published MLPerf offline result. A footnote says the test was by AMD Performance Labs in October 2026, compared with published MLPerf Inference v6.1 results for the Nvidia Vera Rubin NVL72 in the Closed Division, retrieved from mlcommons.org on 11 November 2026, entry 6.1-0106. Those figures are AMD's.
An editor's note says information on the Instinct MI430X was added after the post was first published. The note does not give that later date, so this account does not use the MI430X figures.
What it means for business: the new option named on the post is a 288 GB MI450X, described as matching the MI455X on compute, memory bandwidth and connectivity. The 21,333 tokens per second and the comparison with Nvidia are AMD's own test against a published MLPerf entry. This describes the announcement. It is not advice on a cluster purchase.